Few-shot detection and recognition of thorax diseases in chest x-ray images

In the healthcare domain, some pathologies may be rare and therefore images in the training set may be scarce. Furthermore, traditional supervised machine learning techniques require significantly large datasets which, in a clinical setting, is often laborious to obtain as it necessitates specialist knowledge. Few-shot learning is a sub-area of machine learning and implies that we aim to learn from new data when we have (several classes with) only a few training samples with supervised information. This project is going to address the aforementioned issues in the existing approaches to detection and recognition of thorax diseases in chest x-ray images.

The objectives of this project include:
1) To investigate and implement novel methods to classify thorax diseases from chest x-rays in a few-shot learning scenario.
2) To investigate and implement novel methods to detect/localise thorax diseases from chest x-rays in a few-shot learning scenario.
3) To carefully develop a benchmark for few-shot learning methods which focus on chest x-rays, which is going to overcome the limitation in a lack of appropriate datasets for few-shot learning scenarios.
4) To investigate the effects and impact of base-dataset selection on few-shot learning methods. It has been shown that the selection of the base-dataset considerably influences elements such as the accuracy of the method. We want to verify and measure this effect on our developed methods, as well as examine potential root causes.
5)To establish an effective prototype for few-shot detection and recognition of thorax diseases in chest x-ray images.
The approaches that will be taken to meet the corresponding objectives are:
- For objectives 1) and 2), study the existing few-shot learning methods and learn from their main ideas/frameworks to develop and implement our method which targets chest x-rays.
- For objective 3), examine strategies and guidelines to develop a benchmark using publicly available chest x-ray image datasets.
- For 4) and 5): Use both our own and third-party methods to inspect the effects of base-dataset selection and develop enabling techniques for real applications.